nsingh2

Could you provide some details, if possible, like what model & thinking effort, what kinds of tasks? I used to swap between Claude Code and Codex often, and these days use Codex more because of the usage limits. Wondering if I should go to Claude for a month, I get a strange FOMO when I read vague comments like this.

The one major difference I noticed is that the GPT models are more analytical (e.g. better at mathematical analysis, code review) vs Claude models tend to write more straight forward code. Besides that I don't really see any significant differences.

There are a few gotchas with swapping, like being careful with AGENTS.md/CLAUDE.md naming (Claude Code only recognizes CLAUDE.md, and I think Codex only works with AGENTS.md), and updating skill files to match the tool.

Yup, I meant to write quantization there.

This isn't about training on the output tokens from Anthropic models, it's just about using their models to build things like pretraining pipelines, etc. Even if you train on your own data.

From the phrasing, it might as well be that any ML or infra. related work that even incidentally looks like it could be used to train LLMs may trigger a silent nerf.

It's such an obviously bad policy, it's mind-boggling that they thought this was a good idea. It just breeds paranoia and mistrust, especially when people are already a bit paranoid about silent model quantification for cost cutting reasons.

Works for me, Firefox 151.0 on Linux

Proof search isn't new, but I don't think that captures the value of LLMs.

They act as a learned proposal mechanism on top of hard search. Things like suggesting relevant lemmas, tactics, turning intent into formal steps, and ranking branches based on trained knowledge.

Maybe a kind of learned "intuition engine", from a large corpus of mathematical text, that still has to pass a formal checker. This is not really something we've had to this extent before.

> They do not think

That claim seems less useful, unless “think” is defined in a way that predicts some difference in capability. If the objection is that LLMs are not conscious, fine, but that doesn't say much about whether they can help produce correct formal proofs.

I've been using GPT-5.4, and more recently 5.5, with Codex CLI + Ghidra MCP for reverse engineering a game without many issues. Injecting code is where it usually balks at, but I'm just trying to discover and parse structures from game memory.

I did get a refusal when trying to read in-game currency, even though modifying it would do nothing. It has some strange boundaries.

> a business that puts employees first and profits for owners last can often have a shit ton of profits for owners.

Owners can make 100x that shit ton if they put profits for owners first, so why wouldn’t they do that instead? Out of the goodness of their own hearts?

> The solution, if there is one, has to come from innovation from the private economy.

Why? The problems of offshoring, consolidation, automation, you described came from private sector incentives (not to mention debt driven consumption, and turning basics like housing, healthcare, and education into profit centers)

Why would those same incentives magically fix the problem on their own?

> And there isn't too much the US government can do to revert this economic decline

This is ahistorical. The post great depression economy that led to the “American Dream” was supported by huge public spending and actions by the government [1]. Revitalization happened before, it can happen again.

So much came from FDR/New Deal, social security, labor law, housing finance, banking regulation, securities regulation. Saying the US government can't really do that much is ridiculous.

[1] https://www.archives.gov/milestone-documents/president-frank...

These plots are terrible. Why is categorical data connected across categories with lines? Why not just use bar plots?

Like in the "Web Vulns in OSS" plot, white box data for Opus 4.7 is not available, but the absurd linear interpolation across categories implies it should be near 60.

It's a combination of factors. There was rate-limiting implemented by Anthropic, where the 5hr usage limit would be burned through faster at peak hours, I was personally bitten by this multiple times before one guy from Anthropic announced it publicly via twitter, terrible communication. It wasn't small either, ~15 minutes of work ended up burning the entire 5hr limit. That annoyed me enough to switched to Codex for the month at that point.

Now people are saying the model response quality went down, I can't vouch for that since I wasn't using Claude Code, but I don't think this many people saying the same thing is total noise though.

> https://www.nature.com/articles/s41586-025-09797-z

That title reeks of the paper equivalent of clickbait. The paper is about subjective well-being and mental health in the psychological sense. Broader well-being includes material conditions like income, housing, health care, safety, and social connections. So a null result on subjective well-being is not necessarily a null result on material welfare, and the problems that leads to. The paper’s own abstract also talks about context effects rather than a simple universal null.

> https://marginalrevolution.com/marginalrevolution/2022/10/ar...

Unions are not perfect, but they have been an important check on exploitation. Organized labor helped win the 40-hour workweek. If you demand perfect solutions, you end up doing nothing. And given that you're up against people with nearly unlimited resources, you can't afford to be picky.

This kind of sentiment, on its own, is hollow. Just more "violence bad", until the next round.

There is growing anger and discontentment in a large part of the population, driven by inequality of wealth and power. Hopelessness and a lack of control over the future.

Are the nodes of power willing to spread wealth and control more widely to stabilize the country? What are they willing to do to consolidate their power? The vast majority of violence is perpetrated by those nodes, to either consolidate power, or gain more of it.

Other people in this thread have already suggested more actionable responses: organize, unionize, understand class dynamics, and vote accordingly.

It's going to be expensive to serve (also not generally available), considering they said it's the largest model they've ever trained.

I suspect it's going to be used to train/distill lighter models. The exciting part for me is the improvement in those lighter models.

> They were years ahead.

Considering how fast competitors caught up to them, I'm not convinced that OpenAI was years ahead. LLMs and transformers were known technology, it's just that OpenAI accidentally productized it before others did (ChatGPT). This is not an advantage measured in years. Google, for example, could have caught up to them pretty easily (they invented the transformer architecture), I think it mostly came down to mismanagement that they flopped so hard with Bard. The biggest cost was high quality data, Google certainly had that, and a budget for huge training runs. I really don't think OpenAI had any special sauce that made them years ahead.

One confounder here is that LLM scaling has started to hit diminished returns recently, no more GPT3 -> GPT4/o1 jumps in recent times, making it easier to catch up to the SOTA.

That schism within the OpenAI leadership was ugly. And Sam Altman does seem to be a bit snakey to me. But I have no illusions about any company in this space, including Antropic. None of these companies are moral, given what data these models are trained on.

> their competitors caught up and then passed them by

The different models are more capable in different aspects, but they are close enough together that only in a few months they leapfrog each other.

> OpenAI is floundering and can't sustain their own burn rate. Their competitors are thriving.

Google is thriving, sure, but not because of Gemini, it's because of their existing ads business. I would not say that about Anthropic, they seem to be struggling to provide enough compute (with the recent usage limit changes). Hard to know whats happening funding wise in these companies. Saying that their competitors are thriving is a stretch. And again, if the AI bubble pops, Antropic is gonna hurt along with OpenAI. Just not clear to what extent.

Seems like HN is doing something to combat this, considering how many [dead] comments I see in every post (which you can enable by setting `showdead` in your user profile).

I've only recently enabled it so I don't know how frequent dead comments were before the LLM era.

Really uncharitable take. I did stupid things at 14, and had more unrestricted internet access too.

> absent parent more concerned with his business than his son

I don't know how you came to this conclusion from the post.

This a big exaggeration. Codex is probably one of the top two LLM programming tools, along with Claude Code. GPT-5.4 models are strong, unlike the initial GPT-5 ones, which were comparatively bad, and can hold up against Opus 4.6. In my experience, they are better at analytical work.

I cannot really see how they are "far behind," or how some plugin for Claude Code is a "last desperate bid." The tools are close enough to each other that I regularly use Codex one month and Claude Code the next without much disruption, just to try out any new models or features that might be available.

I do not have much visibility into the non-code applications, so maybe it is stickier there.

If/when the AI bubble pops and takes OpenAI down with it, I would not expect Anthropic to come out unscathed either.

Gemini CLI has been broken for the past 2-3 days, with no response from Google. Really embarrassing for a multi-trillion dollar company. At this point Codex is the only reliable CLI app, out of the big three.

https://www.reddit.com/r/GeminiCLI/comments/1s49pag/this_is_...

This morning I hit 100% 5hr usage on a task that took ~10% in the past. Looks like they are still testing the limits, but it seems over-tuned to me.

Also not great that they communicate this now, since people have been complaining about sudden and strange usage spikes for a few days with no response from Anthropic.

>> More free time?

> Yes! Time we can reclaim from the mundane chores of life to do with as we choose! How could you not want that?

We already had a huge productivity boom these past decades, but wages flat-lined and the vast majority of the profits and surplus went to the top. Housing, education, and healthcare became less affordable, not more. History points against your simple view.

I'm not convinced that AI breaks that pattern. If anything, the concentration is worse this time. The capital required is huge, the technology is controlled by a handful of companies, and the most applications are about replacing labor. That last part further erodes the already meager worker bargaining power.

We do need a serious systemic change to get to the world you're envisioning. One where that congealed wealth needs to start flowing again.

The majority of Ask/Debug mode can be reproduced using skills. For copying code references, if you're using VS Code, you can look at plugins like [1], or even make your own.

Cursor's auto mode is flaky because you don't know which model they're routing you to, and it could be a smaller, worse model.

It's hard to see why paying a middleman for access to models would be cheaper than going directly to the model providers. I was a heavy Cursor user, and I've completely switched to Codex CLI or Claude Code. I don't have to deal with an older, potentially buggier version of VS Code, and I also have the option of not using VS Code at all.

One nice thing about Cursor is its code and documentation embedding. I don't know how much code embedding really helps, but documentation embedding is useful.

[1] https://marketplace.visualstudio.com/items?itemName=ezforo.c...

From [1] (2022 numbers), the median creator earned around 50 Robux per year, which is ~19 cents with the current DevEx rate, and the average was 13,500 Robux.

Out of ~7.5 million creators in 2022, only 11,000 qualified for cashing out.

The distribution is brutal, realistically you have to stick with it for years before getting a hit, if ever. Not to mention the stats probably look worse in the LLM era. You definitely have to like doing it as a hobby.

One caveat is that the creator total likely includes a lot of casual experimentation. If many users make one or two games and then stop (I can see most kids doing this), the 7.5 million figure may overstate how many people are seriously trying to make money from it.

[1] https://about.roblox.com/newsroom/2023/07/vision-roblox-econ...

From what I've read online it's not necessarily a unquantized version, it seems to go through longer reasoning traces and runs multiple reasoning traces at once. Probably overkill for most tasks.

> greed is normal and expected in a free market economy

OK, technically true, just like saying "water flows downhill" when someone's house is flooding. It isn't productive, the fact is well known.

"The system incentivizes this" and "this is good/bad" are two entirely different statements. One doesn't address the other [1], until you make a moral judgement about the outcome.

> You say this as if it's some deviant behaviour that needs correcting.

Is it moral and correct for infants to be fed contaminated baby formula? The mismatch between what is and what ought to be is deviance.

[1] https://en.wikipedia.org/wiki/Is%E2%80%93ought_problem

Most of the value of Claude Code comes from the model, and that's not running on your device.

The Claude Code TUI itself is a front end, and should not be taking 3-4 seconds to load. That kind of loading time is around what VSCode takes on my machine, and VSCode is a full blown editor.

This stuff is a little messy and opaque, but the performance of the same model in different harnesses depends a lot on how context is managed. The last time I tried Copilot, it performed markedly worse for similar tasks compared to Claude Code. I suspect that Copilot was being very aggressive in compressing context to save on token cost, but I'm not 100% certain about this.

Also note that with Claude models, Copilot might allocate a different number of thinking tokens compared to Claude Code.

Things may have changed now compared to when I tried it out, these tools are in constant flux. In general I've found that harnesses created by the model providers (OpenAI/Codex CLI, Anthropic/Claude Code, Google/Gemini CLI) tend to be better than generalist harnesses (cheaper too, since you're not paying a middleman).

In my own testing these models sill have a different flavor to them

- Opus 4.5 for software development. Works faster, and tends to write cleaner code.

- GPT 5.2 xHigh for mathematical analysis, and analysis in general (e.g. code review, planning, double checks), it's very meticulous.

- Gemini 3.0 Pro for image understanding, though this one I haven't played around with much.

There is an argument to be made that that companies like Walmart and Amazon operate as planned economies. They use the same cybernetic principles, real time data monitoring and feedback loops, to solve logistics and planning. These implementations do give credibility Beer's ideas.

There is even a section about this in the wiki article:

https://en.wikipedia.org/wiki/Project_Cybersyn#Contemporary_...

https://en.wikipedia.org/wiki/The_People%27s_Republic_of_Wal...

> Let freedom ring

What happens if the Venezuelan people decide they want their oil profits to stay in Venezuela rather than flowing into oil company coffers? Will they have the "freedom" to choose that?

Don't get me wrong, Maduro being toppled is a positive in isolation, but it's still wait-and-see regarding what he gets replaced with.

"We’re going to have our very large US oil companies, the biggest anywhere in the world, go in, spend billions of dollars, fix the badly broken infrastructure, the oil infrastructure, and start making money for the country and we are ready to stage a second and much larger attack if we need to do so" [1]

[1] https://www.theguardian.com/us-news/2026/jan/03/trump-venezu...